Attention boosted Individualized Regression

Authors: Guang Yang, Yuan Cao, Long Feng

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Comprehensive numerical experiments and real brain MRI analysis using an ADNI dataset demonstrated the superior performance of our model.
Researcher Affiliation Academia Guang Yang Department of Data Science College of Computing City University of Hong Kong guang.yang@my.city.edu.hk Yuan Cao Department of Statistics and Actuarial Science School of Computing and Data Science The University of Hong Kong yuancao@hku.hk Long Feng Department of Statistics and Actuarial Science School of Computing and Data Science The University of Hong Kong lfeng@hku.hk
Pseudocode Yes The pseudocode of the alternating minimization algorithm is summarized in Algorithm 1.
Open Source Code Yes Codes of our approach are available at https://github.com/YLKnight/AIR.
Open Datasets Yes real brain MRI analysis using an ADNI dataset, covering AD patients, individuals with mild cognitive impairment (MCI), and healthy controls. We collected a total of 1059 subjects from ADNI 1 and GO/2 phases.
Dataset Splits Yes The size of the images is set to 28 28, with a sample size of 4000 for training and 1000 for testing. We compare the methods described in the simulation section by 5-fold cross-validation in test RMSE.
Hardware Specification Yes In all experiments, the CPUs used are Intel Xeon Gold 5218R and GPUs used are NVIDIA Ge Force RTX 3090.
Software Dependencies No The paper states that AIR is 'implemented in Python' and ViT is trained using 'Pytorch'. It also mentions 'Matlab code' for LRMR and TRLasso. However, it does not provide specific version numbers for any of these software components or their libraries.
Experiment Setup Yes The AIR is implemented in Python with hyperparameters λ1 and λ2 selected by 5-fold cross-validation, of which the candidate sets are both from 1 to 10. The Vi T is trained by Adam optimizer in Pytorch. Followed by an MLP for regression, the transformer model includes 4 transformer blocks with 8 heads in each Multi-head Attention layer, and the patch size is set as 4 4.